Extractive Summarization of Development Emails

singular vote of each participant, that is already included in the sum of them stored at column "hmn_score") that
could foul up the results. We started the environment with the following settings:
• We chose as attribute evaluator ClassifierSubsetEval 15 that uses a classifier to estimate the "merit" of a set of
attributes.
• The search method was BestFirst “which explores a graph by expanding the most promising node chosen according
to a specified rule" 16 .
• We set "class" as feature to evaluate.
Figure 11. Weka processing platform
Weka Explorer rewarded us with a relevance tree [Figure 12] stating that the 6 attributes which determine the
relevance of a sentence are chars, num_nouns_norm, num_stopw_norm, num_verbs_norm, rel_pos_norm, and
subj_words_norm. Thanks to the tree we could determine whether a sentence should be included in the summary
or not, simply by considering its values of these 6 attributes and going through the conditions written in the tree
(starting from the root). If we arrived to a leaf labeled as "relevant", we include the sentence in the summary,
otherwise not.
15 http://bio.informatics.indiana.edu/ml _docs/weka/weka.attributeSelection.ClassifierSubsetEval.html
16 http://en.wikipedia.org/wiki/Best-first _search
20